reachy_wave / README.md
smilegeng's picture
Use per-episode TsFile layout matching source data_path
004ac31 verified
|
Raw
History Blame
3.81 kB
---
license: apache-2.0
task_categories:
- robotics
- time-series-forecasting
language:
- en
tags:
- lerobot
- robotics
- reachy2
- imitation-learning
- action-state
- video
- tsfile
pretty_name: reachy_wave TsFile
---
# reachy_wave
This repository contains a TsFile conversion of the original Hugging Face
dataset [`aliberts/reachy_wave`](https://huggingface.co/datasets/aliberts/reachy_wave).
## Source Dataset
The source repository is a LeRobot v2.0 dataset for a Reachy2 robot. The task is:
> When you see someone, look at them and wave
Source structure:
```text
data/chunk-000/episode_000000.parquet
data/chunk-000/episode_000001.parquet
data/chunk-000/episode_000002.parquet
data/chunk-000/episode_000003.parquet
data/chunk-000/episode_000004.parquet
meta/episodes.jsonl
meta/info.json
meta/stats.json
meta/tasks.jsonl
videos/chunk-000/observation.images.head_left/episode_000000.mp4
...
videos/chunk-000/observation.images.head_left/episode_000004.mp4
```
Source metadata:
- Robot type: `reachy2`
- Episodes: 5
- Frames: 1,380
- Tasks: 1
- FPS: 15
- Videos: 5
- Video key: `observation.images.head_left`
## Converted Structure
```text
data/chunk-000/episode_000000.tsfile
data/chunk-000/episode_000001.tsfile
data/chunk-000/episode_000002.tsfile
data/chunk-000/episode_000003.tsfile
data/chunk-000/episode_000004.tsfile
meta/episodes.jsonl
meta/info.json
meta/stats.json
meta/tasks.jsonl
videos/chunk-000/observation.images.head_left/episode_000000.mp4
...
videos/chunk-000/observation.images.head_left/episode_000004.mp4
```
The `meta/` and `videos/` directories are mirrored unchanged from the source
repository for traceability. The TsFile conversion uses only the Parquet data.
## Source Schema and Conversion
The source Parquet rows contain:
| Column | Source type | Conversion |
|---|---|---|
| `action` | float32[22] | flattened to `action_0` ... `action_21` |
| `observation.state` | float32[22] | flattened to `state_0` ... `state_21` |
| `timestamp` | float32 seconds | preserved as `episode_timestamp_s`; also used for TsFile `Time` |
| `frame_index` | int64 | FIELD |
| `episode_index` | int64 | FIELD and source for `episode_id` TAG |
| `index` | int64 | renamed to `sample_index` FIELD |
| `task_index` | int64 | FIELD and source for `task_id` TAG |
TsFile details:
- Output files: `data/chunk-000/episode_000000.tsfile` through `data/chunk-000/episode_000004.tsfile`
- Table name: `reachy_wave`
- TAG columns: `episode_id`, `task_id`
- Time precision: milliseconds
- `Time`: `round(timestamp * 1000)`
- Row count preserved: 1,380 source rows -> 1,380 TsFile rows across 5 per-episode files
## Feature Names
Both `action` and `observation.state` use the same 22 dimensions from the source
metadata:
```text
0 neck_yaw.pos
1 neck_pitch.pos
2 neck_roll.pos
3 r_shoulder_pitch.pos
4 r_shoulder_roll.pos
5 r_elbow_yaw.pos
6 r_elbow_pitch.pos
7 r_wrist_roll.pos
8 r_wrist_pitch.pos
9 r_wrist_yaw.pos
10 r_gripper.pos
11 l_shoulder_pitch.pos
12 l_shoulder_roll.pos
13 l_elbow_yaw.pos
14 l_elbow_pitch.pos
15 l_wrist_roll.pos
16 l_wrist_pitch.pos
17 l_wrist_yaw.pos
18 l_gripper.pos
19 mobile_base.vx
20 mobile_base.vy
21 mobile_base.vtheta
```
## Viewer Note
Hugging Face Dataset Viewer may not preview `.tsfile` files because the hosted
viewer environment must have TsFile support installed. Download the per-episode
TsFile files under `data/chunk-000/` and read them with a TsFile-compatible
reader.
## Citation
Please cite the original dataset:
```bibtex
@dataset{reachy_wave,
author = {aliberts},
title = {reachy_wave},
publisher = {Hugging Face},
url = {https://huggingface.co/datasets/aliberts/reachy_wave}
}
```